LiDAR-based lane marking extraction through intensity thresholding and deep learning approaches: a pavement-based assessment

This study presents a normalized intensity thresholding and a deep learning approach with automated labeling for extracting lane markings from LiDAR data. The deep learning model with automated labels achieved the highest accuracy (F1-score: 84.9% in asphalt, 85.1% in concrete), outperforming traditional methods. Results highlight the limitations of intensity-based thresholding on concrete pavements and the effectiveness of deep learning across different surfaces, aiding in up-to-date HD map generation for autonomous vehicles.

Recommended citation: Patel, A., et al. (2020). LiDAR-based Lane Marking Extraction through Intensity Thresholding and Deep Learning Approaches: A Pavement-based Assessment. ISPRS Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download Paper